Motion Synthesis Using Untrained CNN Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current motion synthesis technologies lack the ability to effectively update synthesized motion data based on content and style motion data using untrained convolutional neural networks, limiting their capability in generating accurate feature values for animation applications.
Innovation Solution
A motion synthesis apparatus and method that utilize an untrained convolutional neural network to obtain feature values from content and style motion data, generate target feature values, recognize synthesized motion data, and update the data using a back-propagation algorithm until the synthesized motion feature values match the target feature values, incorporating style loss with weighted assignments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current motion synthesis technologies are used, then the synthesis process can be performed, but the ability to effectively update synthesized motion data based on content and style motion data is lacking
Solution Approach 1:
The patent implements a feedback mechanism where the synthesized motion data is evaluated against target feature values, and the synthesis process is iteratively updated based on the difference (loss) between actual and target features. This feedback loop enables continuous refinement of synthesized motion data to improve accuracy.
Solution Approach 2:
The patent extracts feature values from content and style motion data in advance before the actual synthesis process. These pre-extracted features are then used as targets to guide the synthesis, allowing the system to work with processed information rather than raw data, improving efficiency and accuracy.
2Measurement precision
If untrained convolutional neural networks are used to obtain feature values, then the system can process motion data, but the capability to generate accurate feature values is limited
Solution Approach 1:
The patent extracts feature values from content and style motion data before the synthesis process using an untrained CNN. By performing this feature extraction in advance and using the extracted features as targets for synthesis, the system avoids the need for time-consuming training during the actual synthesis operation.
Solution Approach 2:
The patent creates target feature values by combining content features and style features. These target features serve as a template or copy that the synthesis process aims to reproduce, allowing the system to work toward a known goal without requiring the CNN to learn from scratch during synthesis.
3Manufacturing precision
If iterative updating using back-propagation is implemented, then the accuracy of synthesized motion feature values improves, but the processing time increases
Solution Approach 1:
The patent performs feature extraction from content and style motion data before the iterative synthesis process. This preliminary processing prepares the data in advance, allowing the back-propagation updates to work with pre-processed features rather than raw data, reducing the computational burden during iteration.
Solution Approach 2:
The patent implements continuous iterative updating of synthesized motion data through back-propagation, where each iteration refines the synthesis based on the difference from target features. This continuous refinement process maintains improving precision while the system works efficiently toward convergence.
Data Source
AI summary
A motion synthesis motion synthesis method including: obtaining, by a motion synthesis apparatus, content feature values and style feature values according to content motion data and style motion data; generating, by the motion synthesis apparatus, target feature values using the obtained content feature values and style feature values; recognizing, by the motion synthesis apparatus, synthesized motion data and obtaining synthesized motion feature values from the recognized synthesized motion data; and obtaining, by the motion synthesis apparatus, loss by using the synthesized motion feature values and the target feature values and updating the synthesized motion data according to the obtained loss.